τ-bench

Overview

τ-bench (Tau Bench) is a benchmark for evaluating conversational AI agents that interact with users through domain-specific API tools and policy guidelines. It simulates dynamic, multi-turn conversations where a language model acts as both the user and the agent.

Task Description

Task Type: Conversational Agent Evaluation
Input: User scenarios with specific goals and constraints
Output: Agent actions via API tool calls to complete tasks
Domains: Airline customer service, Retail customer service

Key Features

Dynamic conversation simulation with LLM-simulated users
Domain-specific API tools and policy guidelines
Realistic customer service scenarios
Tests multi-turn dialogue capabilities
Evaluates tool use and policy compliance

Evaluation Notes

Installation Required: pip install git+https://github.com/sierra-research/tau-bench
User Model Configuration: Requires setting up a user simulation model
Primary metric: Accuracy based on task completion reward
Supports airline and retail domains
Uses pass@k aggregation for robustness evaluation
Usage Example

Properties

Property	Value
Benchmark Name	`tau_bench`
Dataset ID	tau-bench
Paper	N/A
Tags	`Agent`, `FunctionCalling`, `Reasoning`
Metrics	N/A
Default Shots	0-shot
Evaluation Split	`test`
Aggregation	`mean_and_pass_hat_k`

Data Statistics

Statistics not available.

Sample Example

Sample example not available.

Prompt Template

No prompt template defined.

Extra Parameters

Parameter	Type	Default	Description
`user_model`	`str`	`qwen-plus`	Model used to simulate the user in the environment.
`api_key`	`str`	`EMPTY`	API key for the user model backend.
`api_base`	`str`	`https://dashscope.aliyuncs.com/compatible-mode/v1`	Base URL for the user model API requests.
`generation_config`	`dict`	`{'temperature': 0.0}`	Default generation config for user model simulation.

Usage

Using CLI

evalscope eval \
    --model YOUR_MODEL \
    --api-url OPENAI_API_COMPAT_URL \
    --api-key EMPTY_TOKEN \
    --datasets tau_bench \
    --limit 10  # Remove this line for formal evaluation

Using Python

from evalscope import run_task
from evalscope.config import TaskConfig

task_cfg = TaskConfig(
    model='YOUR_MODEL',
    api_url='OPENAI_API_COMPAT_URL',
    api_key='EMPTY_TOKEN',
    datasets=['tau_bench'],
    dataset_args={
        'tau_bench': {
            # extra_params: {}  # uses default extra parameters
        }
    },
    limit=10,  # Remove this line for formal evaluation
)

run_task(task_cfg=task_cfg)

2.9 KiB Raw Blame History